AI Agents in Healthcare: What They Do and Where They Fit
AI agents in healthcare are software systems that work toward a defined goal across several steps. They read patient data, take permitted actions in connected systems, and hand the decision back to a person when they reach the limit of their authority. That last part matters more than any feature list.
The word “agent” now gets attached to almost any product that produces text. A note summarizer or a search box over medical records is useful, but it doesn’t plan, act or track whether a task got done. A real agent does all three, and it also knows when to stop.
This guide explains how agents differ from chatbots and scribes, where they already work in clinics and health systems, and what to ask before letting one touch a patient record.
Agents, chatbots, scribes and copilots
The difference comes down to how much authority the system holds. Each type below has a place, and the safest choice is the one with the least authority the job needs.
| System | What it does | Who acts |
|---|---|---|
| Chatbot | Answers one question at a time | The user |
| Scribe | Documents a visit without deciding anything | The clinician |
| Copilot | Suggests options while a person works | The person using it |
| Agent | Pursues a bounded goal over several steps and updates records | The agent, within set permissions |
Inside the agent category there are two working styles. Rules-based agents give the same output for the same inputs, which suits scheduling logic, eligibility checks and reminders. Language-model agents handle open-ended conversation and messy notes, but their answers can vary. Most production systems combine both: fixed rules decide what the agent may do, and the language model works only inside that boundary.
Where AI agents already work
The strongest deployments today sit in front-office and between-visit work, where tasks are repetitive and every step can be checked.
Intake and triage
An intake agent replaces the static form with a conversation. It asks follow-up questions based on earlier answers, flags red-flag symptoms for a clinician, and routes the patient to the right appointment type. Concurrence builds agents for this step that write the intake notes straight into Epic, Oracle Cerner, athenahealth or MEDITECH, and it reports cutting manual intake time by more than 80%.
Scheduling and call handling
Phone lines are where many health systems lose patients first. Hyro runs voice, chat and SMS agents that book, reschedule and answer routine questions. At Intermountain Health, its agents automated 44% of repetitive calls and cut call abandonment by 85%.
Pre-visit preparation
Before a complex follow-up, an agent can gather new lab results, wearable data and questionnaire answers into one sourced brief. It separates what was measured from what the patient reported and lists conflicts for the clinician to resolve, rather than picking a winner itself.
Between-visit follow-up
Agents check in after discharge or a medication change, remind patients about open tasks and escalate when an answer looks wrong. A missed task gets flagged as missed, not read as non-adherence.
Safety escalation
The best agents are tested on edge cases before launch. Concurrence trains its agents against simulated patients first, and at The Care Clinic, a 34-location behavioral health practice, its scheduling agent stops booking and routes the caller to the 988 crisis line when someone discloses a crisis.
Coverage, claims and patients on the move
Payer-side work follows the same pattern: gather documents, check them against rules, and pass anything uncertain to a reviewer. Prior authorization, referral packets and claims intake all fit, because each step has a clear input and a clear owner.
Travel adds a harder version of the same problem. When a patient falls ill abroad, their travel insurance provider usually asks for recent records and a summary of existing conditions before approving treatment. An agent can find the right documents, prepare the request and track the reply. Releasing those records still needs the patient’s consent and a staff member’s sign-off.
Travel also changes the data itself. A patient who just crossed several time zones will show disrupted sleep and heart-rate readings. A good agent notes the trip next to the data, so the clinician doesn’t read jet lag as a new problem.
Six checks before you deploy
A demo shows what an agent does when things go right. These checks show what it does when they don’t.
- Data. Which systems does it read, and does every statement trace back to a source? Conflicts and gaps should be visible, not smoothed over.
- Goal. “Improve outcomes” is not a goal. “Prepare a sourced visit brief and never propose treatment” is.
- Permissions. Reading, drafting, queuing, sending and changing a care plan are separate rights. Diagnosis and treatment stay with clinicians.
- Tools. For each connected system, ask whether the agent reads or writes, whether the action can be undone, and who hears about a failure.
- Stop rules. Contradictory doses, undated labs, crisis disclosures and out-of-scope requests should all halt the agent and reach a person.
- Ownership. “Human in the loop” is too vague. Name who reviews, who approves, and where the decision is recorded.
One warning from the research: a widely used population-health algorithm once used healthcare spending as a proxy for need, and it underrated Black patients’ illness as a result (Obermeyer et al., 2019). The goal you give an agent shapes who it serves.
Where this goes next
The next step is not more autonomy but tighter hand-offs. Fixed rules set what an agent may do, the language model works inside that line, automatic checks confirm each change, and a named person approves anything that matters. The vendors worth trusting are the ones who can show the exact moment their agent hands the decision back.
FAQ
Are AI agents in healthcare allowed to diagnose patients? No, not in most current deployments. Agents prepare information, route tasks and flag risks, while diagnosis and treatment decisions stay with licensed clinicians.
How is an AI agent different from a medical chatbot? A chatbot answers one question at a time. An agent works toward a goal across several steps and can update records within set permissions.
What should a clinic ask in a vendor demo? Ask to see the agent decide to do nothing, handle conflicting data, show its sources, and fail safely when a connected system goes down.
Ivy Joy
Helping to build Mazurly from the ground up, managing content, operations, digital communication, everything from resource development and customer relationships to strategic partnerships and platform growth.